To briefly summarize the issues reported by syzbot, there are two task call stacks
as follows:
Task A Task B
------------------------------------ ----------------------------------
drm_atomic_commit
drm_atomic_helper_commit
commit_work
commit_tail
drm_atomic_helper_commit_tail commit_tail
drm_atomic_helper_commit_tail
drm_client_modeset_commit_atomic
drm_atomic_state_default_clear
drm_atomic_helper_wait_for_vblanks
When two prerequisites are met simultaneously, the current issue will be triggered:
1. There is an overlap in the memory range occupied by the crtcs member set contained
in the instance state of "struct drm_atomic_state" created by Task A and Task B
2. The context of drm_atomic_helper_commit_tail() has no lock protection, resulting
in the instance state->crtcs sub item being released by other task
The solution is to add a lock in drm_atomic_helper_commit_tail() to ensure that
there is no other task interference when accessing the instance state.
Reported-and-tested-by: syzbot+0f999d26a4fd79c3a23b@xxxxxxxxxxxxxxxxxxxxxxxxx
Signed-off-by: Edward Adam Davis <eadavis@xxxxxx>
---
drivers/gpu/drm/drm_atomic_helper.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
index 39ef0a6addeb..b16ff9020097 100644
--- a/drivers/gpu/drm/drm_atomic_helper.c
+++ b/drivers/gpu/drm/drm_atomic_helper.c
@@ -1743,7 +1743,9 @@ EXPORT_SYMBOL(drm_atomic_helper_wait_for_flip_done);
void drm_atomic_helper_commit_tail(struct drm_atomic_state *old_state)
{
struct drm_device *dev = old_state->dev;
+ static DEFINE_MUTEX(lock);
+ mutex_lock(&lock);
drm_atomic_helper_commit_modeset_disables(dev, old_state);
drm_atomic_helper_commit_planes(dev, old_state, 0);
@@ -1757,6 +1759,7 @@ void drm_atomic_helper_commit_tail(struct drm_atomic_state *old_state)
drm_atomic_helper_wait_for_vblanks(dev, old_state);
drm_atomic_helper_cleanup_planes(dev, old_state);
+ mutex_unlock(&lock);
}
EXPORT_SYMBOL(drm_atomic_helper_commit_tail);
--
2.43.0
